magistrsko delo
Abstract
V magistrskem delu je predstavljena metoda iskanja utežem agnostičnih nevronskih mrež, ki temelji na genetskem algoritmu, imenovanem NeuroEvolution of Augmenting Topologies (NEAT). Evalviranje genomov z vzorčenjem uteži iz fiksne uniformne množice naključnih vrednosti minimizira pomembnost uteži, s čimer je poudarek le na optimizaciji topologije. To omogoča utežem agnostičnim nevronskim mrežam opravljanje različnih nalog brez predhodnega učenja utežnih vrednosti. Naša implementacija je bila prilagojena za povezovanje z odprtokodno knjižnico Scikit-learn, ki smo jo javno objavili v obliki PyPi paketa. V eksperimentalnem delu smo se osredotočili na primerjavo evolucijskih in utežem agnostičnih nevronskih mrež na primeru reševanja klasifikacijskih problemov. Rezultate smo evalvirali z uporabo statističnih metod, ki so pokazale, da utežem agnostične nevronske mreže proizvedejo več skritih nevronov kot evolucijske, vendar uspejo doseči primerljivo točnost zgolj s pravilno topologijo, brez optimizacije uteži.
Keywords
utežem agnostične nevronske mreže;klasifikacija;nevroevolucija;magistrske naloge;
Data
Language: |
Slovenian |
Year of publishing: |
2020 |
Typology: |
2.09 - Master's Thesis |
Organization: |
UM FERI - Faculty of Electrical Engineering and Computer Science |
Publisher: |
[M. Mlakar] |
UDC: |
004.8(043.2) |
COBISS: |
45029891
|
Views: |
400 |
Downloads: |
76 |
Average score: |
0 (0 votes) |
Metadata: |
|
Other data
Secondary language: |
English |
Secondary title: |
Classification with weight agnostic neural networks |
Secondary abstract: |
In our master's thesis, we reviewed a search method for weight agnostic neural networks that are based on a genetic algorithm called NeuroEvolution of Augmenting Topologies (NEAT). Evaluating genomes by sampling weights from a fixed uniform random distribution ensures the importance of weights is minimized and the main focus is on optimizing the topology. This gives weight agnostic neural networks an ability to solve different tasks without explicit weight training. Our implementation was made to be compatible with an open-source library called Scikit-learn, and we published it as a public PyPi package. In our experiments, we focused on comparing evolutionary neural networks with weight agnostic neural networks by solving different classification tasks. We evaluated the results with the use of statistical methods which showed that while weight agnostic neural networks created more hidden nodes, their topologies were able to achieve comparable accuracy without optimizing the weights. |
Secondary keywords: |
weight agnostic neural networks;classification;neuroevolution;NEAT; |
Type (COBISS): |
Master's thesis/paper |
Thesis comment: |
Univ. v Mariboru, Fak. za elektrotehniko, računalništvo in informatiko, Informatika in tehnologije komuniciranja |
Pages: |
XI, 73 f. |
ID: |
12111121 |